Deep Learning and Traditional Machine Learning for Automated Driver Distraction Detection

The development of intelligent in-vehicle systems capable of automatically detecting driver inattentiveness is critical for enhancing road safety and reducing traffic accidents. This paper details a comprehensive comparative study evaluating a modern deep learning approach against traditional machine learning classifiers for the task of driver distraction detection. The primary model is a custom-designed Convolutional Neural Network (CNN), implemented in Python using the Keras library, trained to classify driver behavior from dashboard camera images into ten distinct states. In contrast to classical machine learning models that require manual feature engineering, the CNN learns a hierarchy of spatial features directly from raw image data. To provide a rigorous performance benchmark, the CNN is evaluated against three traditional algorithms: Random Forest, Gaussian Naive Bayes, and a Linear Support Vector Classifier (SVC). All models were trained and tested on a large, publicly available dataset. The system architecture, data preprocessing pipeline, and model designs are described in detail. Our experimental results conclusively demonstrate the superior performance of the CNN, which achieved a validation accuracy exceeding 98%. This work demonstrates that for vision-based classification tasks, the end-to-end feature learning capability of deep learning models provides an unequivocal advantage over traditional methods that rely on pre-processed, flattened feature vectors.